Improving Logical-level Natural Language Generation with Topic-conditioned Data Augmentation and Logical Form Generation

Logical Natural Language Generation, i.e., generating textual descriptions that can be logically entailed by a table, has been challenge due to the low fidelity of generation. Previous works have addressed this problem annotating interim logical programs control generation contents and semantics, presented task table-aware form text (Logic2text) However, although table instances are abundant in real world, forms paired with require costly human labor, which limits parallel data size. To mitigate this, we propose topic-conditioned augmentation (TopicDA), employs controlled sequence-to-sequence as auxiliary tasks augment directly from tables. We further introduce (LG), dual Logic2text requires valid based on description table. hence semi-supervised learning approach jointly train an LG model both labeled augmented data. Experimental results demonstrate our effectively utilize outperform supervised baselines substantial margin.